Researchers often point to two gaps between current AI and general intelligence.
Continual Learning
People learn continuously: each experience updates what they know. Most AI models are trained once, then deployed with fixed weights. They can use information in their context window, but don't permanently learn from it.
Challenges include:
- Catastrophic forgetting: training on new data can degrade old abilities.
- Cost: retraining large models is expensive.
- Safety: models that change after deployment are harder to evaluate.
Current workarounds include retrieval, memory systems and periodic fine-tuning.
Generalisation
Models perform well on problems similar to their training data. Performance can drop on genuinely novel situations or unusual variants of familiar problems. Debate continues over how much models truly reason versus pattern-match, and the answer may differ by task.
Why They Matter
A general intelligence should adapt to new jobs, tools and environments without being rebuilt. Progress on these capabilities would mark a significant step.
Research Directions
New architectures, training methods, memory systems and ways to learn efficiently from few examples are all under active research.